Cascaded boosted predictive models

Abstract

Cascaded, boosted predictive models trained using distinct sets of exogenous and endogenous features are configured to predict component of performance ratings of entities. From the distinct predicted components, the second entity's rating factor can be determined. A second entity's rating factor represents the specific contribution a second entity makes to his average performance rating, as distinct from the rating that an arbitrary or hypothetical second entity would obtain.

Description

CROSS REFERENCE TO RELATED APPLICATIONS[0001]This application claims the benefit of U.S. Provisional Application No. 62 / 307,368, filed Mar. 11, 2016, which is hereby incorporated by reference in its entirety.TECHNICAL FIELD[0002]This disclosure generally relates to the field of statistical machine learning in which systems have the capability to automatically update a current set of data and relationships using data mining, and more particularly to applications of boosting classifiers.[0003]Fields of Classification: Artificial Intelligence (Class 706), subclass 12 (Machine Learning); 707 (Data Processing), subclasses 776 (Data Mining); Cross Reference Art Collection: 905, 913 (Vehicles).BACKGROUND[0004]Statistical machine learning is a class of computer-implemented data analysis algorithms that identify relationships between variables (equivalently features or attributes) in complex data sets, for applications in classification, clustering, regression (prediction), ranking, and othe...

AI Technical Summary

Benefits of technology

[0008]Unlike conventional gradient boosting methods, in which all of the models are trained on the same set of features (even accounting for random feature selection in random forest methods), in the present system, each model is trained on different category of features that are distinct from the specific features used in the other models, and thus the models selectively predict the distinctive contribution each category of features contributes to the performance rating, providing improved feature and model decorrelation with respect to conventional gradient boosted approaches. This leads to improved explanatory power of the models in terms of being able to identify the most significant variables in each category of features.

[0015]The ROAR factor for a given driver can then be used in various examples. One use of the ROAR factor is to select the drivers who will receive a given trip request: drivers with high ROAR factors are more likely than the average driver in the current circumstances to obtain a maximal performance rating from the rider—which is beneficial to increase the likelihood of the rider having a positive experience, and thus the trip request is provided by the transportation management system to candidate drivers who have a ROAR factor above a threshold value or range. Another use of the ROAR factor is to selectively provide messages to the driver before, during or after a trip that instruct behavioral changes in the driver so as to increase the likelihood of the driver receiving a higher performance rating score. Another use of the ROAR factor is to identify drivers who are likely not meet performance rating goals in the future and provide them with additional training to improve their performance.

Problems solved by technology

For example, existing methods do not inherently take into account whether an independent feature occurs before, during, or after an event or value being classified or predicted, whether a feature is an aspect of the environment itself or the result of an actor or agent in that environment.

Similarly, existing methods do not distinguish between features that are spontaneously observed and those that are intentionally elicited, or between features that assumed and those that are experimental.

These limitations result in limited ability of existing approaches to use these various types of contextual features of the data to improve the accuracy of the model.

Because such a rating reflects the rider's overall experience, it may, in fact, be based upon factors that are beyond the control of the driver—for example, a rider may provide a poor rating to the driver for a trip made during rush hour since the rider was delayed in reaching their destination, even though the rush hour traffic was beyond the control of the driver.

Because there are numerous factors that can contribute to the rating given by the rider, conventional approaches that strictly look at a driver's average performance rating do not accurately assess the driver's actual contribution to the performance rating.

As such, existing ensemble learning methods do not immediately lend themselves to applications for evaluating driver performance in transportation management systems.

Images